In-process semantic search evaluating unit-normalized dense embeddings directly via dot products in memory. Sub-millisecond latency, zero network API calls, and zero token costs.